Which action can make the little man jump up to the hightest green pip in the first world?
Bojue-Wang opened this issue · comments
when i train an agent, even after the little man can jump over most of the first world, it can be frequently stucked under those highest green pip, and try a lot of actions, it seems the succeed to jump up on top of the pip is by accident.
When i play a mario with a gamepad, it seems i can control the height with the duration i push the jump button.
what is the mechanism to transfer long-jump-push to the action in this gym framekork?
The mechanism for the agent is exactly the same as for a human. To produce a long jump your agent has to emit a sequence of "right + A" actions or something similar over time. The problem you're noticing is either due to your learning algorithm or model being unable to learn the dynamics of the environment.